David Chiang presents a statistical machine translation model that uses hierarchical phrases, which are phrases containing subphrases. The model is formally a synchronous context-free grammar (CFG) learned from parallel text without syntactic annotations, combining ideas from syntax-based and phrase-based translation. The system's training and decoding methods are detailed, and it is evaluated for translation speed and accuracy using BLEU. The system outperforms the Alignment Template System, a state-of-the-art phrase-based system. The introduction discusses the limitations of phrase-based models and the benefits of hierarchical phrases. The grammar extraction process is described, including the use of word alignments and constraints to filter and refine the grammar. The model is defined using a log-linear approach, incorporating features such as lexical weights and penalties for different rule types. The training process involves estimating parameters using relative-frequency estimation and minimum-error-rate training. The decoding process uses a CKY parser with beam search and a postprocessor for mapping French derivations to English translations. The system's performance is evaluated using BLEU, and the implementation details are provided.David Chiang presents a statistical machine translation model that uses hierarchical phrases, which are phrases containing subphrases. The model is formally a synchronous context-free grammar (CFG) learned from parallel text without syntactic annotations, combining ideas from syntax-based and phrase-based translation. The system's training and decoding methods are detailed, and it is evaluated for translation speed and accuracy using BLEU. The system outperforms the Alignment Template System, a state-of-the-art phrase-based system. The introduction discusses the limitations of phrase-based models and the benefits of hierarchical phrases. The grammar extraction process is described, including the use of word alignments and constraints to filter and refine the grammar. The model is defined using a log-linear approach, incorporating features such as lexical weights and penalties for different rule types. The training process involves estimating parameters using relative-frequency estimation and minimum-error-rate training. The decoding process uses a CKY parser with beam search and a postprocessor for mapping French derivations to English translations. The system's performance is evaluated using BLEU, and the implementation details are provided.